Benchmarking Self-Supervised Contrastive Learning Methods for Image-based Plant Phenotyping
dc.contributor.committeeMember | Eramian, Mark | |
dc.contributor.committeeMember | Stavness, Ian | |
dc.contributor.committeeMember | Neufeld, Eric | |
dc.contributor.committeeMember | Mondal, Debajyoti | |
dc.creator | Ogidi, Franklin C. | |
dc.creator.orcid | 0000-0002-6386-4784 | |
dc.date.accessioned | 2023-01-19T14:49:48Z | |
dc.date.available | 2023-01-19T14:49:48Z | |
dc.date.copyright | 2022 | |
dc.date.created | 2023-01 | |
dc.date.issued | 2023-01-19 | |
dc.date.submitted | January 2023 | |
dc.date.updated | 2023-01-19T14:49:48Z | |
dc.description.abstract | Image-based plant phenotyping enables the high-throughput measurement of the physical characteristics of plants by combining one or more imaging technologies with image analysis tools. Over the past decade, deep learning has been widely successful for image-based tasks like image classification, object detection, image segmentation and object counting. While deep learning has been applied to image-based plant phenotyping tasks like plant species classification, plant disease detection, and leaf counting, its application has been limited. Part of the reason for this is that deep learning models tend to rely on large annotated datasets for training, and it can be expensive and time consuming to generate such datasets. Motivated by the need to leverage unlabelled data, a lot of research effort has recently been directed towards the area of self-supervised learning (SSL). The common theme among various SSL methods is that they derive the supervisory signal from the data itself, usually by distorting the input in some way and learning features that are invariant to the distortions. Despite the surge of research in this area, there has been a paucity of research applying self-supervised learning on image-based plant phenotyping tasks, particularly detection and counting tasks. We address this gap by benchmarking two self-supervised learning methods -- MoCo v2 and DenseCL -- on four image-based plant phenotyping tasks (the downstream tasks): wheat head detection, plant instance detection, wheat spikelet counting and leaf counting. We study the effects of the domain of the pre-training dataset on the transfer performance using four large-scale datasets: ImageNet (general purpose concepts), iNaturalist 2021 (natural world images), iNaturalist 2021 Plants (plant images) and the TerraByte Field Crop datatset (crop images). To understand the differences between the internal representations of the neural networks trained with the different methods, we applied a representation similarity analysis technique known as orthogonal Procrustes distance. Our results show that (1) Finetuning a model that is pre-trained with an SSL method typically outperforms training from scratch for a downstream task, (2) The Supervised pre-training method outperforms DenseCL and MoCo v2 for all the downstream tasks, except for the leaf counting task where DenseCL excels, (3) There is not much difference, both in the downstream performance and the internal representations, between MoCo v2 and DenseCL pre-trained models, (4) Pre-training with the iNaturalist 2021 Plants dataset leads to the best downstream performance more often than other datasets, and (5) Models pre-trained in a supervised manner learn more dissimilar features towards the last layers compared to models pre-trained with MoCo v2 or DenseCL. We hope that this benchmark/evaluation study will inspire further studies towards the development of better self-supervised representation learning methods for image-based plant phenotyping tasks. | |
dc.format.mimetype | application/pdf | |
dc.identifier.uri | https://hdl.handle.net/10388/14428 | |
dc.language.iso | en | |
dc.subject | self-supervised learning | |
dc.subject | plant phenotyping | |
dc.subject | contrastive learning | |
dc.subject | representation similarity analysis | |
dc.subject | wheat head detection | |
dc.subject | global wheat head detection | |
dc.subject | wheat spikelet counting | |
dc.subject | leaf counting | |
dc.subject | plant detection | |
dc.subject | MoCo | |
dc.subject | DenseCL | |
dc.title | Benchmarking Self-Supervised Contrastive Learning Methods for Image-based Plant Phenotyping | |
dc.type | Thesis | |
dc.type.material | text | |
thesis.degree.department | Computer Science | |
thesis.degree.discipline | Computer Science | |
thesis.degree.grantor | University of Saskatchewan | |
thesis.degree.level | Masters | |
thesis.degree.name | Master of Science (M.Sc.) |